MRI and CT images are the most popular formats assisting a doctor in diagnosis and treatment, but highly accurate segmentation is a challenging problem due to intensity inhomogeneity and environmental noises. In this paper, we introduce an appropriate and effective automatic approach to facilitate this problem in two stages. In the first stage, skull region is removed from the brain by morphological active contour and level set process. Moreover, in level set process, some AI rules are defined on slice positions of brain to increase the accuracy. In the second stage, a modified EM method is performed on the resultant skull-stripping image to identify some candidate main regions of CSF (cerebro-spinal fluid), GM (gray matter), and WM (white matter). The candidate regions are then re-estimated into the proper CSF, GM, and WM through a Bayesian Estimation Process. The experimental results show that the proposed approach obtains a robust segmentation for IBSR, OASIS and Korean Hospital database. With the proposed AI-rules, the level set method gets good skull-stripping images regardless of MRI slice position in bran. Also, Bayesian postprocessing can improve the segmentation performance by 10~15% in CSF, GM and WM ratios compared the basic EM algorithm.